Are you ready for the digital AIpocalypse?

Mano Paul opens by arguing that people do not share a common definition of AI, which makes it hard to defend. He separates narrow, general and super intelligence, notes that most corporate AI is predictive rather than generative, and points to heavy overlap between the OWASP ML Top 10 and LLM Top 10.

Next he walks through attack categories. Data poisoning corrupts training data so a model misclassifies, illustrated with a medical dosing example. Prompt injection, both direct and indirect, can make an LLM application ignore its instructions or take actions it was never meant to take, tied to excessive agency. Model inversion can infer sensitive attributes such as age from outputs, while model poisoning alters training parameters, often through an insider or a weak supply chain. Model evasion adds crafted perturbations to inputs, and a live demo applies that noise to a tiger shark photo for an image classifier.

He closes with practical notes: traditional controls such as access control to models and datasets still apply, and teams need ongoing monitoring for drift, performance metrics and prompt logs that someone actually reviews. For newcomers he suggests starting with Python and notebooks, since understanding how to break software helps when building it.